LSTM Network Based On Adaptive Prediction Model For Person Re-Identification

With the rapid development of video surveillance system in the public safety field,the research on person re-identification has become more and more in-depth.Person re-identification refers to giving a target person under a certain camera,and then re-identify the target person under the remaining cameras through an algorithm.However,the performance of person re-identification algorithm is easily affected by factors such as scale,illumination,viewing angle,pose transformation,and occlusion.These factors make it difficult for person re-identification to be widely used in security.Therefore,the development of advanced person re-identification algorithms and techniques has resulted in Important significance.Compared with single image,the video sequence contains more temporal information that can be used to improve person re-identification performance.However,how to model temporal information in person re-identification is a challenging problem.The LSTM(Long Shot Term Memory)network can easily remember the long-term interdependence of sequence data.Zhang et al.[32]proposed an image-to-video person re-identification method based on convolutional neural network(CNN)and LSTM in 2017.This method has achieved very good performance.Although the LSTM network can accumulate sequence information of video,one of the defects is that the output of the LSTM is more biased towards the data after the time step in the network.In the actual surveillance video,person may be in the situation where they may walk into the shield or out of the shield.Using the LSTM model when the person walk into the shield half will cause the LSTM feature of the target person contains the information of after entering the shield;when the person out of the shield half,Using the LSTM model will cause the LSTM feature of the target person contains the information of before leaving the shield.The LSTM features generated in these two cases will affect the performance of person re-identification,To overcome the above situation,this paper proposes an Adaptive Prediction Model Selection Network(APM-Net)for person without shield,walk into the shield half and out of the shield half are used for prediction.The forward LSTM and backward LSTM network characteristics are adaptively selected through the prediction mode.This improves the feature distinguishability and improves the recognition performance.This method is mainly composed of two parts:feature extraction and distance metric learning.The feature extraction part includes:?Performing LBP&Color feature extraction on each frame in the continuous video sequence;?Inputting the LBP&Color features mentioned in the continuous video sequence as input into the forward and backward LSTM networks to obtain the forward LSTM and the backward LSTM Output characteristics;?through the APM-Net network to obtain the forward and backward LSTM features adaptive selection to get video frame characteristics.The distance measurement learning section takes two sets of person's video features as input and determines whether the corresponding person is same through the XQDA distance measurement learning.The method was tested on two standard pedestrian data sets iLIDS-VID and PRID 2011.For the data set iLIDS-VID with occlusion,the rankl value of this method is 54.93%,which is 4.53%higher than the RFA-Net method and 3.2%higher than the bidirectional LSTM feature method.The experimental results show that the proposed pedestrian recognition method based on adaptive prediction mode for LSTM network has good re-identification performance for occlusion.